| CVE |
Vendors |
Products |
Updated |
CVSS v3.1 |
| Unbounded memory accumulation in the stream message reassembly path of Apache bRPC (version < 1.18.0) on all platforms allows remote attackers to exhaust the memory of a server (denial of service) by sending an endless sequence of continuation data frames.
Root Cause:
The streaming RPC protocol bounds every single data frame by max_body_size, but a fragmented stream message accumulates continuation frames into one buffer (_pending_buf) without any limit on the accumulated size. A malicious peer that can open a stream can therefore keep sending continuation frames to make the receiver accumulate an arbitrarily large in-memory message. Since the default flow-control settings (socket_max_streams_unconsumed_bytes=0) do not throttle the sending rate, the receiver's memory grows until the process runs out of memory and crashes or becomes unavailable.
Affected Scenarios:
Services that accept streams from untrusted clients (for example using brpc::StreamAccept or brpc::StreamCreate with the built-in streaming protocol) are affected.
How to Fix:
(Choose one of the following options)
1. Upgrade bRPC to version 1.18.0, which fixes this issue.
2. Apply this patch: https://github.com/apache/brpc/pull/3489
Note:
The fix limits the accumulated payload of a fragmented stream message to max_body_size (default 64MB). When the accumulated payload exceeds the limit, the pending fragments are released and the stream is closed with error EMSGSIZE. After applying the fix, legitimate fragmented stream messages larger than max_body_size will be rejected; you can adjust the gflag max_body_size to change the limit. |
| vLLM is an inference and serving engine for large language models. From 0.16.0 until 0.32.0, the /v1/responses endpoint allows an authenticated caller to place an attacker-controlled template in chat_template_kwargs[chat_template] when no server template is configured. ChatParams.get_apply_chat_template_kwargs() preserves that value without enforcing trust_request_chat_template, so the request can bypass the intended trust policy and cause CPU or memory exhaustion through Jinja rendering. This issue is fixed in version 0.32.0. |
| vLLM is an inference and serving engine for large language models. From 0.19.0 until 0.32.0, authenticated callers can supply a request chat template when no server-side --chat-template is configured, or a malicious model repository can provide a template that is compiled and rendered by tokenizer.apply_chat_template() without a wall-clock timeout, output-size limit, or iteration bound. Rendering runs in a ThreadPoolExecutor with one worker by default, so a nonterminating render can block /v1/responses and other chat-template users after the request ends, while an output-expanding render can exhaust server memory. Harmony models are not affected because they do not use this Jinja path. This issue is fixed in version 0.32.0. |
| Twisted is an event-based framework for internet applications, supporting Python 3.6+. Prior to 26.10.0rc1, twisted.words.protocols.irc.IRC.dataReceived retains all bytes received without a line feed in its per-connection buffer and applies no maximum line length. The twisted.words.service.IRCUser server subclass inherits this behavior, and buffering occurs before PASS, NICK, or USER authentication commands. An unauthenticated connection can continuously grow the buffer until the shared server process runs out of memory and drops all users. This issue is fixed in version 26.10.0rc1. |
| toml-node is a TOML parser for Node.js and the browser. Prior to 5.0.1, toml.parse() passes unbounded hexadecimal, octal, binary, and decimal digit sequences from lib/parser.js to BigInt() before reduceInteger() in lib/compiler.js checks whether the resulting value fits the TOML 64-bit integer range. Converting a sufficiently long integer string consumes super-linear CPU time and synchronously blocks the Node.js event loop, allowing one attacker-supplied TOML document to deny service to concurrent work in the process. The later range check cannot prevent the conversion cost because it executes after BigInt(). This issue is fixed in version 5.0.1. |
| In the Linux kernel, the following vulnerability has been resolved:
drm/amdgpu: fix lock leak on ENOMEM in AMDGPU_GEM_OP_GET_MAPPING_INFO
The AMDGPU_GEM_OP_GET_MAPPING_INFO branch of amdgpu_gem_op_ioctl()
holds three cleanup-tracked resources before calling kvcalloc():
the drm_gem_object reference from drm_gem_object_lookup(), the
drm_exec lock on the looked-up GEM via drm_exec_lock_obj(), and
the drm_exec lock on the per-process VM root page directory via
amdgpu_vm_lock_pd(). All three are released by the out_exec
label that every other error path in this function jumps to.
The kvcalloc() failure path returns -ENOMEM directly, skipping
out_exec and leaking all three.
The leaked per-process VM root PD dma_resv lock is the
load-bearing leak: any subsequent operation on the same VM
(further GEM ops, command-submission, eviction, TTM shrinker
callbacks) blocks on the held lock. DRM_IOCTL_AMDGPU_GEM_OP is
DRM_AUTH | DRM_RENDER_ALLOW, so this is an unprivileged-local
denial of service against the caller's GPU context, reachable
by any process with /dev/dri/renderD* access.
Route the failure through out_exec so drm_exec_fini() and
drm_gem_object_put() run.
Reproduced on stock 7.0.0-10, Ryzen 7 5700U / Radeon Vega
(Lucienne): the failing ioctl returns -ENOMEM and a second
GET_MAPPING_INFO on the same fd then blocks in
drm_exec_lock_obj() on the leaked dma_resv. SIGKILL on the
caller does not reap the task; the fd-release path during
process exit goes through amdgpu_gem_object_close() ->
drm_exec_prepare_obj() on the same lock, leaving the task in D
state until the box is rebooted. The patched kernel was not
rebuilt and re-tested on this hardware; the fix is mechanical.
Tested on a single Lucienne / Vega box only.
Ziyi Guo posted an independent INT_MAX-bound check for
args->num_entries in the same branch [1]; the two patches are
complementary and can land in either order.
(cherry picked from commit b69d3256d79de15f54c322986ff4da68f1d65b0a) |
| Starlet versions from 0.20 before 0.33 for Perl allow memory exhaustion via unbounded buffering when decoding chunked request bodies.
The decoder adds each read to one buffer of unconsumed input. Nothing limits the length of a chunk-size line, so a line without a CRLF grows the buffer for as long as the client keeps sending, and the read timeout restarts after every read. The data of a large chunk is also held in the buffer until the whole chunk has arrived.
A remote attacker can exhaust a worker's memory with a single request. |
| In the Linux kernel, the following vulnerability has been resolved:
mm/migrate_device: fix pgtable leak in migrate_vma_insert_huge_pmd_page
When migrate_vma_insert_huge_pmd_page() jumps to unlock_abort due
to a PMD check failure, the pgtable allocated earlier via
pte_alloc_one() is never freed, causing a memory leak.
Added free_abort label to release the pgtable in error path. |
| In the Linux kernel, the following vulnerability has been resolved:
iio: buffer: Fix DMA fence leak in iio_buffer_enqueue_dmabuf()
iio_buffer_enqueue_dmabuf() allocates a struct iio_dma_fence (104 bytes,
kmalloc-128) via kmalloc_obj()+dma_fence_init(), which sets the initial
kref to 1. It then calls dma_resv_add_fence() which takes a second
reference (kref=2), and stores a raw pointer in block->fence.
On the success path the function returns without calling dma_fence_put()
to release the initial reference, so every buffer enqueue permanently
leaks one kmalloc-128 allocation.
The iio_buffer_cleanup() work item only releases the temporary reference
taken during completion signalling by iio_buffer_signal_dmabuf_done();
the initial reference from dma_fence_init() is never released.
With four iio_rwdev instances at 240kHz and 512 samples per buffer,
this produces ~1875 kmalloc-128 allocations per second matching the
observed slab growth exactly. A test with ftrace confirmed that the
dma_fence_destroy event was never triggered.
Fix by calling dma_fence_put() after dma_resv_add_fence(), transferring
ownership of the fence to the DMA reservation object. The DMA fence then
gets properly discarded after being signalled. |
| The Redis protocol implementation in Apache bRPC allowed inline command parsing to allocate and copy the full connection buffer before a complete CRLF-terminated command was available. A remote attacker with access to a Redis-protocol bRPC service could send long inline inputs without a terminating CRLF to increase memory usage and CPU copy overhead, potentially causing denial of service.
Root Cause:
The Redis inline command parser allocated memory based on the current input buffer size and copied the entire buffer before determining the current inline command boundary. The existing redis_max_allocation_size limit was applied to RESP bulk strings and arrays, but not effectively to the inline command path before allocation.
Affected Scenarios:
Use Apache bRPC server with Redis protocol enabled and reachable by untrusted clients.
How to Fix:
1. Upgrade Apache bRPC to version 1.18.0, which fixes this issue.
2. Apply the patch from PR https://github.com/apache/brpc/pull/3363 .
Note:
The fix applies redis_max_allocation_size to Redis inline command lines and scans for CRLF before allocating parser arena memory. Inline command lines exceeding this limit are rejected. |
| The RTMP AMF parser in Apache bRPC accepts AMF string lengths and array counts directly from untrusted input. A remote unauthenticated attacker who can reach a bRPC server with RTMP enabled can send a small AMF payload with an excessive declared long string length or array count, causing excessive memory allocation or resource consumption before the parser rejects the incomplete payload.
Root Cause:
The AMF long string parser resized the destination string to the declared 32-bit length before validating that the stream contained that many bytes. ECMA array and strict array parsers also accepted 32-bit element counts without a configurable upper bound.
Affected Scenarios:
Deployments that enable the bRPC RTMP server component and accept RTMP traffic from untrusted clients. Direct uses of the AMF parser on untrusted data are also affected.
How to Fix:
1. Upgrade bRPC to version 1.18.0, which fixes this issue.
2. Apply this patch: https://github.com/apache/brpc/pull/3353
Note:
The fix introduces configurable AMF parsing limits. The default maximum string size is 64 MiB and the default maximum array element count is 1,048,576. Payloads above these limits are rejected. The limits can be adjusted with amf_max_string_size and amf_max_array_size. |
| Apache Kafka's code-generated readers for flexible protocol versions read a tagged-field count declared by the sender and retain a RawTaggedField object for each unknown field. That count was bounded only by the outer request frame size (socket.request.max.bytes); it was not checked against the remaining buffer bytes or capped. Because each entry can be encoded in as little as two bytes (unknown tag 0, size 0), a modest request can declare tens of millions of tagged fields and force one object allocation per declared field. Flexible request headers (header version 2, used by flexible APIs such as ApiVersions v3) are parsed this way, so the amplification applies to the request header itself.
Impact:
The flexible request header is parsed on the network thread before authorization, request dispatch, and request-time quota accounting. A client can send a request of a few tens of megabytes whose header declares about 20 million tagged fields, forcing roughly that many RawTaggedField allocations and exhausting broker heap or inducing process-wide GC stalls, denying broker service. Because the header is parsed for pre-authentication APIs such as ApiVersions (v3), the allocation can occur before authentication; no Kafka resource ACL or quota is required.
Mitigation:
Upgrade to Apache Kafka 4.2.2, 4.3.2, or 4.4.0. Until then, lowering socket.request.max.bytes and restricting network access to broker listeners reduces, but does not remove, exposure.
Resolution:
From Apache Kafka 4.2.2, 4.3.2, and 4.4.0: generated readers bounds-check the tagged-field count for flexible-versioned messages, rejecting a negative count, a count exceeding the remaining buffer bytes, or a count exceeding a fixed maximum (10,000), before any per-field allocation. |
| The nsenter --join-cgroup option opens the target cgroup.procs file as root and leaves that file descriptor open across later namespace and credential changes and across execve(). Because the kernel checks later cgroup migrations using the credentials from the original open, a program run in an attacker-controlled target can inherit root's ability to move host processes between cgroups. After a privileged operator uses --join-cgroup against that target, an unprivileged user can migrate and terminate unrelated root processes. |
| IBM Langflow OSS 1.0.0 through 1.12.4 could allow a remote authenticated attacker to cause a denial of service due to allocation of resources without limits or throttling. |
| Crypt::Age versions before 0.004 for Perl allow CPU exhaustion via an unbounded number of recipient stanzas, each costing an X25519 scalar multiplication in unwrap_file_key.
The header parser accepts any number of recipient stanzas, and unwrap_file_key tries every identity against every X25519 stanza. Each attempt runs a scalar multiplication, an HKDF and a ChaCha20-Poly1305 open. The header MAC cannot be checked first, because its key comes from the file key, which only a successful unwrap produces.
Each stanza is 98 bytes of attacker-chosen header and buys one scalar multiplication per identity held, and the file is rejected only after every one of them has run. A caller reaches this through decrypt, decrypt_file or decrypt_filehandle whenever it decrypts a file it did not write. |
| DBI versions before 1.656 for Perl hang on a later open of a DBD::DBM table because a failed open_data does not release the table's lock.
open_data creates the table's lock file and takes an exclusive flock on it for a CREATE, before the checks that can croak: a dbm_mldbm value that is not a module name, a failing tie, or a missing data file. The croak unwinds before the table object is blessed, so nothing releases the lock, and the meta data holding the open descriptor stays cached on the database handle.
The dbm_mldbm and dbm_type values come from the DSN or the connect attributes, so a caller that chooses the serializer or the storage backend can trigger the failure.
The flock has no timeout, so a later open of that table (even through a separate database handle) blocks for as long as the handle holding the stale lock lives, in the same process or any other. |
| In the Linux kernel, the following vulnerability has been resolved:
gpiolib: fix memory leak in gpiod_set_array_value_complex() on error paths
A drive-by review by Sashiko reports that, after a failed call to
gpiochip_set_multiple(), gpiod_set_array_value() returns abruptly,
bypassing the freeing of possibly alloced bitmaps, potentially leaking
memory. It is an unlikely error path, but is possible.
Defer the return code check until after the bitmap_free() block. |
| In the Linux kernel, the following vulnerability has been resolved:
xfs: fix capability check in xfs
An user reported a bug where he managed to evade group's quota
by changing a file's gid to a different group id the same user
belonged to, even though quotas were enforced on both gids and the
file's size was big enough to exceed the quota's hardlimit.
Commit eba0549bc7d1 replaced a capable() call by a
has_capability_noaudit() to prevent unnecessary selinux audit messages.
Turns out that both calls have slightly different semantics even though
their documentation seems similar. Where in a nutshell:
capable() - Tests the task's effective credentials
has_ns_capability_noaudit() - Tests the task's real credentials
This most of the time has no practical difference but in some cases like
changing attrs (specifically group id in this case) through a NFS client
this will allow the quota code to use XFS_QMOPT_FORCE_RES, effectively
bypassing quota accounting checks.
Using instead ns_capable_noaudit() should fix this issue and prevent
selinux audit messages.
This also fix the remaining calls to has_capability_noaudit() |
| In the Linux kernel, the following vulnerability has been resolved:
drm/amdgpu: free the userptr ranges when a CS fails after parser_bos
amdgpu_cs_parser_bos() allocates an HMM range for every userptr entry of
the BO list and looks up the user pages. Those ranges are only freed when
amdgpu_cs_parser_bos() itself fails or when amdgpu_cs_submit() checks
them under the notifier lock. If amdgpu_cs_patch_jobs(),
amdgpu_cs_vm_handling(), amdgpu_cs_sync_rings() or the early part of
amdgpu_cs_submit() fails, nothing frees them: neither
amdgpu_cs_parser_fini() nor amdgpu_bo_list_free() touch e->range.
Each such CS then leaks the range, its hmm_pfns array (8 bytes per page
of the userptr BO) and the BO reference the range holds. An IB on a ring
that patches the CS in place (e.g. VCN) with an unmapped va_start fails
in amdgpu_cs_patch_ibs() and takes this path.
Free the ranges on those error paths. Ranges amdgpu_cs_submit() already
dropped are NULL, so freeing them again is a no-op.
(cherry picked from commit c52a9542ed7cfd527feb938bd0accada2ac8d267) |
| In the Linux kernel, the following vulnerability has been resolved:
fs: Free any excess xarray nodes in clear_inode()
For many years we've had a hard to hit leak of xarray nodes. Hugh
documented it well in commit 786b31121a2c. Recently people and
syzbot have found ways to force it to happen with madvise. Rather
than fix the leaks where they happen, just call xa_destroy() which
has the side-effect of cycling the i_pages lock. |
